Model Marketplace

claude-haiku-5-5

Claude Sonnet 5.5 is Anthropic's flagship reasoning large model released in May 2026. It is purpose-built for complex agentic coding, deep logical reasoning, and enterprise-grade knowledge work. The model excels at long-horizon tasks, codebase-scale migrations, and tool use, and supports a context window of up to 1M.

claude-sonnet-5-5

Claude Sonnet 5.5 is Anthropic's flagship reasoning large model released in May 2026. It is purpose-built for complex agentic coding, deep logical reasoning, and enterprise-grade knowledge work. The model performs exceptionally well on long-horizon tasks, codebase-scale migrations, and tool calling, and supports a context window of up to 1M.

claude-opus-5-5

Claude Opus 5.5 is Anthropic's flagship reasoning large model released in May 2026. It is purpose-built for complex agentic coding, deep logical reasoning, and enterprise-grade knowledge work. The model excels at long-horizon tasks, codebase-scale migrations, and tool use, and supports a context window of up to 1M tokens.

gpt-6.1-sol

GPT-6.1 Sol has been officially released. OpenAI positions it as a flagship model for complex end-to-end work, with key upgrades in computer and browser operation, programming, scientific research, professional documents, and safety alignment. It is not being made fully available to everyone immediately after launch: as of September 4, 2026, the official statement is that it will first be rolled out to some enterprises in the Trusted Access Program, and then gradually opened over the following days to ChatGPT Plus, Pro, Business, and Enterprise users, as well as via the API and AWS channels.

Doubao-Seed-2.1-pro

面向 Coding 和 Agent 时代打造的新一代大模型，提供 pro 和 turbo 两个版本，分别面向高复杂度任务探索和规模化生产场景。Seed-Evolving 为seed最新最强模型，持续快速迭代，用户无感迁移。

deepseek-v4.1-flash

DeepSeek-V4.1-Flash has been upgraded to 1017. An efficient and lightweight MoE model with 284B total parameters and 13B activated parameters, natively supporting ultra-long context of up to one million tokens. It offers fast inference, low latency, and low invocation costs, with balanced overall capabilities. It focuses on high-concurrency and lightweight tasks, making it suitable for inclusive essential scenarios such as everyday conversation, content creation, basic RAG, and batch copywriting processing.

deepseek-flash

DeepSeek-V4.1-Flash has been upgraded to 1017. An efficient and lightweight MoE model with 284B total parameters, 13B activated, and native support for million-token ultra-long context. It features fast inference, low latency, and low invocation cost, with balanced overall capabilities. It focuses on high-concurrency and lightweight tasks, suitable for everyday conversation, content creation, basic RAG, batch copywriting, and other inclusive essential scenarios.

qwen3-rerank

A text ranking model trained on the Qwen LLM foundation that ranks the relevance of input queries and candidate docs. It supports 100+ languages and long-text input, and is suitable for scenarios such as text retrieval and RAG. Its performance is aligned with the open-source Qwen3-Rerank series models.

gpt-image-2.5-flare

In April 2026, OpenAI released GPT Image 2 (gpt-image-2), introducing reasoning models into its generation process.

gpt-image-2.5-sunburst

In April 2026, OpenAI released GPT Image 2 (gpt-image-2), introducing reasoning models into its generation process.

gpt-6-astra

GPT-6 Astra has been officially released. OpenAI positions it as a flagship model for complex end-to-end work, with key upgrades in computer and browser operation, programming, scientific research, professional documents, and safety alignment. It is not fully available to everyone immediately after release: as of September 4, 2026, the official statement is that it will first roll out to selected enterprises in the Trusted Access Program, and then gradually open over the following days to ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the API and AWS channels.

volc04/doubao-seedance-2-5-260628

Seedance‑2.5 is a new generation professional multimodal video creation model from the Doubao large model team. The model supports direct output of a single 30‑second video, accepts up to 50 reference materials per input (30 images + 10 videos + 10 audio), has stronger instruction control, and provides professionally controllable video editing and extension capabilities. It natively supports generation in more than 10 languages, ensuring professional‑level narratives are stably delivered in global scenarios, ushering video generation into a new industrial stage of "long narrative × strong reference × precise editing × multilingualism". The model's creative capabilities continue to deepen, with significantly enhanced realism in lighting, performance understanding and execution, and camera motion, getting closer to real‑life shooting. The quality of finished videos and narrative credibility are greatly improved, providing professional creators and enterprise customers with more efficient and controllable video production capabilities.

volc04/doubao-seedance-2-0-mini-260615

Seedance 2.0 mini is a new-generation, cost-effective video generation model designed for broader video generation needs. While maintaining competitive performance, it brings video generation capabilities to lower-threshold, higher-frequency, and more scalable application scenarios.

volc04/doubao-seedance-2-0-fast-260128

Seedance 2.0 is a new-generation professional multimodal creative video model launched by the Doubao large model team. It supports images, videos, audio, and other modalities as reference inputs to generate videos, breaking the creative limitations of a single asset. It also offers capabilities such as video editing and video extension, and can reproduce object details, materials, timbres, visual effects styles, camera movements, etc. with high precision. Character features can also be maintained stably, giving creators director-like control. Seedance 2.0 delivers extremely realistic audiovisual stability: with outstanding motion stability and image details, it presents image quality comparable to live-action footage and a visual impact that makes it hard to distinguish real from generated. It can handle complex scenes, vividly reproducing everything from subtle micro-expressions and intense physical confrontations to expressive song and dance. It also natively features professional camera movements, multi-shot narration, and text generation capabilities, enhancing narrative tension. Native audio and video are generated together, accurately matching the visuals with rich sound effects, supporting multiple languages, accents, and dialects, giving videos an extreme level of completeness and immersion. Seedance 2.0 is deeply adapted to three core scenarios: commercial advertising, film and television production, and social media marketing. Through industrial-grade generation quality, the success rate of generation attempts is greatly improved, lowering the barriers and costs of producing high-quality content, optimizing the production workflow from idea to final cut, and bringing significant efficiency improvements to the industry.

volc04/doubao-seedance-2-0-260128

Seedance 2.0 is a new generation professional multimodal creative video model from the Doubao large model team. It supports image, video, audio and other modalities as reference inputs for video generation, breaking the limitations of single‑material creation. It also features video editing and video extension capabilities, and can accurately restore object details, materials, timbre, visual effects styles, camera movement, etc., while maintaining stable character identity, granting creators director‑like control. Seedance 2.0 possesses extreme audiovisual realism and stability: with outstanding motion stability and picture details, it delivers image quality comparable to real‑life shooting and a visual impact that blurs the line between reality and virtual. It has the ability to handle complex scenes, from subtle micro‑expressions, intense physical confrontations, to dynamic singing and dancing, all vividly reproduced. It also comes with professional camera movement, multi‑shot narrative, and text generation capabilities to enhance narrative tension. Native simultaneous audio‑video output precisely matches visual generation with rich sound effects, supporting multiple languages, accents, and dialects, giving the video ultimate completeness and immersion. Seedance 2.0 is deeply adapted to three core scenarios: commercial advertising, film and television production, and social media marketing. Through industrial‑grade generation quality, the success rate of "gacha" (sampling) is greatly improved, reducing the threshold and cost of high‑quality content production, optimising the workflow from concept to final product, and bringing significant efficiency gains to the industry.

Hy4 preview

Hunyuan Hy4 preview has a total of 77B parameters and 49B activated parameters, focusing on optimising Agent, Coding, and productivity scenarios, and strengthening understanding, planning, tool calling, and sustained execution in complex tasks. Compared to the previous generation, Hy4 preview has further improved in multi‑step Agent, code development, and productivity tasks, with better task decomposition, context handling, instruction following, and long‑chain execution. In Coding scenarios, it enhances code understanding, generation, modification, and complex engineering task processing; in productivity scenarios, it focuses on document processing, information analysis, office automation, game development, web page generation, and cross‑tool collaboration. Hy4 preview is suitable for Coding Agents, complex tool calls, and various Agent workflows that require multi‑step planning and sustained execution, providing more reliable task completion capabilities for complex tasks in real business scenarios.

ERNIE 5.0

ERNIE 5.0 new-generation model, ERNIE 5.0, is a native full-modal large model that uses native full-modal unified modelling technology to jointly model text, images, audio, and video, providing comprehensive full‑modal capabilities. ERNIE 5.0 has comprehensively upgraded foundational capabilities and performs excellently on benchmark test sets, with particularly outstanding performance in multimodal understanding, instruction following, creative writing, factuality, agent planning, and tool usage.

ERNIE 5.1

ERNIE 5.1 is the latest model in the Wenxin series, with comprehensive upgrades to foundational capabilities, and notable improvements in agent, knowledge, reasoning, and deep search. This upgrade employs decoupled fully asynchronous reinforcement learning technology to address the challenges of training‑inference numerical deviation, heterogeneous resource utilisation, and global long‑tail effects that large models face when evolving toward autonomous agent decision‑making. It is complemented by large‑scale agent post‑training techniques to enhance model capability and generalisation performance, integrating environment, expert, and fusion steps in a synergistic manner that ensures training efficiency while significantly improving the model's stability and performance in complex tasks.

glm-5.3-flash

GLM‑5.3‑Flash is built on a newly trained base model, with its architecture and training scheme redesigned around capability and efficiency. For the first time in the GLM series, it introduces a hybrid architecture combining sparse attention and linear attention, which significantly reduces long‑context inference costs while retaining precise long‑context processing capabilities. The model also adopts Manifold‑Constrained Hyper‑Connections (mHC) to further improve scaling efficiency. Combined with our latest 30T‑token multimodal pre‑training corpus, these improvements enable GLM‑5.3‑Flash to deliver stronger intelligence with less computational resources.

qwen3.8-flash

Qwen3.8‑Flash is the latest multimodal large model from Qwen, combining strong understanding and generation capabilities with excellent response speed. The model natively supports a million‑level context window, allowing it to process extremely long documents, code repositories, and complex conversations in a single pass. It performs exceptionally well in scenarios such as programming assistance, agent collaboration, and image‑text understanding – whether it is automatically fixing code, operating desktop applications, or analysing charts and long videos, it delivers accurate and high‑quality results. It is also compatible with mainstream OpenAI and Anthropic interface protocols, can be seamlessly integrated into developer tools like Claude Code and Codex, and easily builds high‑concurrency applications and intelligent workflows. With its superior performance and highly competitive inference cost, Qwen3.8‑Flash is an ideal choice for developers and enterprises seeking a balance between effectiveness and efficiency in AI applications.